📄 ch3_progress.c
字号:
mpi_errno = MPI_SUCCESS; }#if 0 if (msgqIter++ == MPIDI_CH3I_MSGQ_ITERATIONS) { msgqIter = 0; /*printf("[%d] calling message queue progress from test.\n", MPIR_Process.comm_world->rank);fflush(stdout);*/ mpi_errno = MPID_CH3I_Message_queue_progress(); if (mpi_errno != MPI_SUCCESS) { mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**mqp_failure", 0); } }#else mpi_errno = MPIDI_CH3I_Message_queue_progress(); if (mpi_errno != MPI_SUCCESS) { mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**mqp_failure", 0); }#endiffn_exit: MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3_PROGRESS_TEST); return mpi_errno;}#undef FUNCNAME#define FUNCNAME MPIDI_CH3_Progress_wait#undef FCNAME#define FCNAME MPIDI_QUOTE(FUNCNAME)int MPIDI_CH3_Progress_wait(MPID_Progress_state *state){ int mpi_errno = MPI_SUCCESS; int rc; MPIDU_Sock_event_t event; unsigned completions = MPIDI_CH3I_progress_completion_count; int num_bytes; MPIDI_VC_t *vc_ptr; /*static int active = 0;*/ static int updateIter = 0; static int msgqIter = MPIDI_CH3I_MSGQ_ITERATIONS; MPID_CPU_Tick_t start, end; static MPID_CPU_Tick_t shmTicks = 0; static int shmIter = 0; static int shmReps = 1; static int shmTotalReps = 0; static int spin_count = 1; static MPID_CPU_Tick_t sockTicks = 0; static int sockIter = 0; static int sockReps = 1; static int sockTotalReps = 0; MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3_PROGRESS_WAIT); /*MPIDI_STATE_DECL(MPID_STATE_MPIDU_YIELD);*/ MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3_PROGRESS_WAIT); MPIDI_DBG_PRINTF((50, FCNAME, "entering")); do { /* make progress on the shared memory queues */ if (sockIter != 0) goto skip_shm_loop; MPID_CPU_TICK(&start); for (; shmIter<shmReps; shmIter++, shmTotalReps++) { if (MPIDI_CH3I_Process.shm_reading_list) { rc = MPIDI_CH3I_SHM_read_progress( MPIDI_CH3I_Process.shm_reading_list, 0, &vc_ptr, &num_bytes); if (rc == MPI_SUCCESS) { MPIDI_DBG_PRINTF((50, FCNAME, "MPIDI_CH3I_SHM_read_progress reported %d bytes read", num_bytes)); mpi_errno = MPIDI_CH3I_Handle_shm_read(vc_ptr, num_bytes); if (mpi_errno != MPI_SUCCESS) { mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**progress", 0); goto fn_exit; } } else { if (rc != SHM_WAIT_TIMEOUT) { /*MPIDI_err_printf("MPIDI_CH3_Progress", "MPIDI_CH3I_SHM_read_progress returned error %d\n", rc);*/ mpi_errno = MPIR_Err_create_code(rc, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**progress", 0); goto fn_exit; } /* if (rc == SHM_WAIT_TIMEOUT_WHILE_ACTIVE) { active = active | MPID_CH3I_SHM_BIT; } */ } if (MPIDI_CH3I_shm_read_active) spin_count = 1; if (completions != MPIDI_CH3I_progress_completion_count) { MPIDI_CH3I_active_flag |= MPID_CH3I_SHM_BIT; /* active = active | MPID_CH3I_SHM_BIT; spin_count = 1; */ shmIter++; shmTotalReps++; goto after_shm_loop; } } if (MPIDI_CH3I_Process.shm_writing_list) { vc_ptr = MPIDI_CH3I_Process.shm_writing_list; while (vc_ptr) { if (vc_ptr->ch.send_active != NULL) { rc = MPIDI_CH3I_SHM_write_progress(vc_ptr); if (rc == MPI_SUCCESS) { /*active = active | MPID_CH3I_SHM_BIT;*/ if (completions != MPIDI_CH3I_progress_completion_count) { MPIDI_CH3I_active_flag |= MPID_CH3I_SHM_BIT; spin_count = 1; shmIter++; shmTotalReps++; goto after_shm_loop; } } else if (rc != SHM_WAIT_TIMEOUT) { mpi_errno = MPIR_Err_create_code(rc, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**progress", 0); goto fn_exit; } } vc_ptr = vc_ptr->ch.shm_next_writer; } }/* if (spin_count++ >= MPIDI_Process.my_pg->ch.nShmWaitSpinCount) { MPIDI_FUNC_ENTER(MPID_STATE_MPIDU_YIELD); MPIDU_Yield(); MPIDI_FUNC_EXIT(MPID_STATE_MPIDU_YIELD); spin_count = 1; }*/ /* FIXME: Why do the other yields have MPIDI_FUNC_ENTER/EXIT macros around them but this one does not? */ MPIDU_Yield(); }after_shm_loop: if (shmIter == shmReps) { shmIter = 0; } MPID_CPU_TICK(&end); shmTicks += end - start; if (shmIter != 0) goto skip_sock_loop;skip_shm_loop: MPID_CPU_TICK(&start); for (; sockIter<sockReps; sockIter++, sockTotalReps++) { /* make progress on the sockets */ mpi_errno = MPIDU_Sock_wait(MPIDI_CH3I_sock_set, 0, &event); if (mpi_errno == MPI_SUCCESS) { mpi_errno = MPIDI_CH3I_Progress_handle_sock_event(&event); if (mpi_errno != MPI_SUCCESS) { mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**progress_sock_wait", 0); goto fn_exit; } /*active = active | MPID_CH3I_SOCK_BIT;*/ MPIDI_CH3I_active_flag |= MPID_CH3I_SOCK_BIT; } else { if (MPIR_ERR_GET_CLASS(mpi_errno) != MPIDU_SOCK_ERR_TIMEOUT) { mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**progress_sock_wait", 0); goto fn_exit; } /* comment out this line to test the error functions */ mpi_errno = MPI_SUCCESS; /* FIXME: Why do the other yields have MPIDI_FUNC_ENTER/EXIT macros around them but this one does not? */ MPIDU_Yield(); } if (completions != MPIDI_CH3I_progress_completion_count) { sockIter++; sockTotalReps++; goto after_sock_loop; } }after_sock_loop: if (sockIter == sockReps) { sockIter = 0; updateIter++; } MPID_CPU_TICK(&end); sockTicks += end - start;skip_sock_loop: if (updateIter == MPIDI_CH3I_UPDATE_ITERATIONS) { updateIter = 0; switch (MPIDI_CH3I_active_flag) { case MPID_CH3I_SHM_BIT: /* only shared memory has been active */ /* give shm MPID_SINGLE_ACTIVE_FACTOR cycles for every 1 sock cycle */ shmReps = (int) ( ( sockTicks * (MPID_CPU_Tick_t)MPID_SINGLE_ACTIVE_FACTOR * (MPID_CPU_Tick_t)shmTotalReps ) / (MPID_CPU_Tick_t)sockTotalReps / shmTicks); if (shmReps < 1) shmReps = 1; if (shmReps > 256) { shmReps = 256; } sockReps = 1; /*MPIU_DBG_PRINTF(("(SHM_BIT: shmReps = %d, sockReps = %d)", shmReps, sockReps));*/ break; case MPID_CH3I_SOCK_BIT: /* only sockets have been active */ /* give sock MPID_SINGLE_ACTIVE_FACTOR cycles for every 1 shm cycle */ sockReps = (int) ( ( shmTicks * (MPID_CPU_Tick_t)MPID_SINGLE_ACTIVE_FACTOR * (MPID_CPU_Tick_t)sockTotalReps ) / (MPID_CPU_Tick_t)shmTotalReps / sockTicks ); if (sockReps < 1) sockReps = 1; if (sockReps > 100) { sockReps = 100; } shmReps = 1; /*MPIU_DBG_PRINTF(("(SOCK_BIT: shmReps = %d, sockReps = %d)", shmReps, sockReps));*/ break; case MPID_CH3I_SHM_BIT + MPID_CH3I_SOCK_BIT: /* both channels have been active */ /* give each channel 50% of the spin cycles */ shmReps = (int) ( ( (MPID_CPU_Tick_t)((shmTicks + sockTicks) >> 1) * (MPID_CPU_Tick_t)shmTotalReps ) / shmTicks ) / MPIDI_CH3I_UPDATE_ITERATIONS; if (shmReps < 1) shmReps = 1; if (shmReps > 256) { shmReps = 256; } sockReps = (int) ( ( (MPID_CPU_Tick_t)((shmTicks + sockTicks) >> 1) * (MPID_CPU_Tick_t)sockTotalReps ) / sockTicks ) / MPIDI_CH3I_UPDATE_ITERATIONS; if (sockReps < 1) sockReps = 1; if (sockReps > 100) { sockReps = 100; } /*MPIU_DBG_PRINTF(("(BOTH_BITS: shmReps = %d, sockReps = %d)", shmReps, sockReps));*/ break; default: /* neither channel has been active */ /* leave the state in its previous state */ /*printf(".");*/ break; } /* if (shmReps > 500 || sockReps > 100) { printf("[%d] SHMREPS=%d, SOCKREPS=%d\n", MPIR_Process.comm_world->rank, shmReps, sockReps); } */ MPIDI_CH3I_active_flag = 0; shmTotalReps = 0; shmTicks = 0; shmIter = 0; sockTotalReps = 0; sockTicks = 0; sockIter = 0; } if (msgqIter++ == MPIDI_CH3I_MSGQ_ITERATIONS) { msgqIter = 0; /*printf("[%d] calling message queue progress\n", MPIR_Process.comm_world->rank);fflush(stdout);*/ /* printf("%d",2*MPIR_Process.comm_world->rank); */ mpi_errno = MPIDI_CH3I_Message_queue_progress(); if (mpi_errno != MPI_SUCCESS) { mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**mqp_failure", 0); goto fn_exit; } /* printf("%d",2*MPIR_Process.comm_world->rank+1); */ } } while (completions == MPIDI_CH3I_progress_completion_count);fn_exit: MPIDI_DBG_PRINTF((50, FCNAME, "exiting, count=%d", MPIDI_CH3I_progress_completion_count - completions)); MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3_PROGRESS_WAIT); return mpi_errno;}#endif /* USE_ADAPTIVE_PROGRESS */#ifdef USE_FIXED_ACTIVE_PROGRESS/**********************************************************//* *//* Fixed active progress engine *//* *//* Like the adaptive engine but active reps are fixed *//* instead of calculated from CPU time. *//* *//**********************************************************/#define MPIDI_CH3I_UPDATE_ITERATIONS 10#define MPID_SINGLE_ACTIVE_FACTOR 100/* FIXME: What is this routine for? Who uses it? Why is it not a static function? */#undef FUNCNAME#define FUNCNAME MPIDI_CH3I_Message_queue_progress#undef FCNAME#define FCNAME MPIDI_QUOTE(FUNCNAME)int MPIDI_CH3I_Message_queue_progress(){ MPIDI_CH3I_Shmem_queue_info info; int num_bytes; MPIDI_VC_t *vc_ptr; int rc, mpi_errno;#if defined(HAVE_SHARED_PROCESS_READ) && !defined(HAVE_WINDOWS_H) char filename[256];#endif /* check for new shmem queue connection requests */ rc = MPIDI_CH3I_BootstrapQ_recv_msg( MPIDI_Process.my_pg->ch.bootstrapQ, &info, sizeof(info), &num_bytes, FALSE); if (rc != MPI_SUCCESS) { mpi_errno = MPIR_Err_create_code(rc, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**boot_recv", 0); return mpi_errno; }#ifdef MPICH_DBG_OUTPUT /*MPIU_Assert(num_bytes == 0 || num_bytes == sizeof(info));*/ if (num_bytes != 0 && num_bytes != sizeof(info)) { mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**bootqmsg", "**bootqmsg %d", num_bytes); return mpi_errno; }#endif if (num_bytes) { MPIDI_PG_t *pg; MPIDI_PG_Find(info.pg_id, &pg); MPIDI_PG_Get_vc(pg, info.pg_rank, &vc_ptr); /*vc_ptr = &MPIDI_Process.my_pg->ch.vc_table[info.pg_rank];*/ rc = MPIDI_CH3I_SHM_Attach_to_mem( &info.info, &vc_ptr->ch.shm_read_queue_info); if (rc != MPI_SUCCESS) { mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**MPIDI_CH3I_SHM_Attach_to_mem", "**MPIDI_CH3I_SHM_Attach_to_mem %d", vc_ptr->ch.shm_read_queue_info.error); /*"MPIDI_CH3I_SHM_Attach_to_mem failed, error %d\n", vc_ptr->ch.shm_read_queue_info.error);*/ return mpi_errno; } MPIU_DBG_PRINTF(("attached to queue from process %d\n", info.pg_rank));#ifdef HAVE_SHARED_PROCESS_READ#ifdef HAVE_WINDOWS_H /*MPIU_DBG_PRINTF(("Opening process[%d]: %d\n", i, pSharedProcess[i].nPid));*/ vc_ptr->ch.hSharedProcessHandle = OpenProcess(STANDARD_RIGHTS_REQUIRED | PROCESS_VM_READ | PROCESS_VM_WRITE | PROCESS_VM_OPERATION, FALSE, info.pid); if (vc_ptr->ch.hSharedProcessHandle == NULL) { int err = GetLastError(); mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**OpenProcess", "**OpenProcess %d %d", info.pg_rank, err); return mpi_errno; }#else MPIU_Snprintf(filename, 256, "/proc/%d/mem", info.pid); vc_ptr->ch.nSharedProcessID = info.pid; vc_ptr->ch.nSharedProcessFileDescriptor = open(filename, O_RDWR/*O_RDONLY*/); if (vc_ptr->ch.nSharedProcessFileDescriptor == -1) { mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**open", "**open %s %d %d", filename, info.pid, errno); return mpi_errno; }#endif#endif /*vc_ptr->ch.state = MPIDI_CH3I_VC_STATE_CONNECTED;*/ /* we are read connected but not write connected */ vc_ptr->vc_ptr->ch.shm_read_connected = 1; vc_ptr->ch.bShm = TRUE; vc_ptr->ch.read_shmq = vc_ptr->ch.shm_read_queue_info.addr; MPIU_DBG_PRINTF(("read_shmq = %p\n", vc_ptr->ch.read_shmq)); vc_ptr->ch.shm_reading_pkt = TRUE; /* add this VC to the global list to be shm_waited on */ MPIDI_CH3I_SHM_Add_to_reader_list(vc_ptr); } return MPI_SUCCESS;}#undef FUNCNAME#define FUNCNAME MPIDI_CH3I_Progress#undef FCNAME#define FCNAME MPIDI_QUOTE(FUNCNAME)int MPIDI_CH3I_Progress(int is_blocking, MPID_Progress_state *state){ int mpi_errno = MPI_SUCCESS; int rc; MPIDU_Sock_event_t event; unsigned completions = MPIDI_CH3I_progress_completion_count; int num_bytes; MPIDI_VC_t *vc_ptr; /*static int active = 0;*/ static int updateIter = 0; static int msgqIter = 0; MPID_CPU_Tick_t start, end; static MPID_CPU_Tick_t shmTicks = 0;
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -